Claude docs changes for September 18th, 2026 [diff]

Executive Summary

  • Projects land in public beta: a 462-line new page for one ongoing conversation where Claude starts a cloud-session thread per task, all sharing the project's repositories, instructions and memory. Pro and Max only, rolling out gradually to accounts that have used cloud sessions
  • Skills and plugins enabled on your claude.ai account now sync into terminal sessions signed in with it (2.1.275), and claude.ai can list marketplaces — such as an organization plugin library — that you add by name with claude plugin marketplace add --claudeai
  • New startup_failure_reason on SDK results, with a table of sixteen values naming exactly why Claude Code refused to start, so an application can offer the fix instead of retrying
  • A send-now key (ctrl+enter) interrupts the current turn and sends all queued messages at once, with sent and queued messages shown in gray until the model receives them
  • /plugin install <plugin> --marketplace <source> adds a marketplace and installs from it in one step, asking you to confirm the resolved source first

New Claude Code versions

2.1.275

New features

  • Added syncing of the skills and plugins enabled on your claude.ai account to terminal sessions signed in with it; opt out with syncClaudeAiSkills: false or syncClaudeAiPlugins: false
  • Added a send-now key (ctrl+enter, or ctrl+x ctrl+s) that interrupts the current turn and sends all queued messages at once
  • Added /plugin install <plugin> --marketplace <source>, which offers to add the marketplace before installing
  • Added the signed-in account to Claude apps gateway sign-in: when the gateway names it, you confirm it before the credential is saved, and /status shows it
  • Added a startup warning when a configured otelHeadersHelper fails, so sessions that silently export no telemetry are noticed

Existing feature improvements

  • Improved prompt caching for a --system-prompt containing a __SYSTEM_PROMPT_DYNAMIC_BOUNDARY__ line: the text above it is now cached globally
  • Improved the Artifact tool's publish and read results to say who can open the page and what the owner's Share menu offers
  • Improved artifact publish results: they name the tab icon sent, warn when the page contains a NUL byte, and retry a flaky fetch of a newer page after a stale publish
  • Improved the /desktop error when Claude Desktop doesn't open

Major bug fixes

  • Fixed a whole family of resume crashes and hangs on malformed transcripts: a malformed task-reminder or @-file attachment entry, a malformed message entry, and a malformed message content block each used to break --resume, the resume picker preview, resumed background agents and the transcript view
  • Fixed /rewind in a forked or background session restoring a zero-filled or truncated file when the session's file-history backups couldn't be fully copied
  • Fixed plugin and marketplace messages, logs and claude plugin marketplace list showing a password or token stored in a git, ssh or marketplace URL
  • Fixed claude plugin marketplace update deleting a GitHub marketplace's local copy when the fetch failed and the marketplace was named after its repository
  • Fixed sandboxed Bash commands on Linux reporting exit code 0 for failed commands when the shell is zsh
  • Fixed sandboxed Bash commands being unable to write to project directories named hooks/ or config/
  • Fixed Grep, Glob and @-file suggestions hanging or running out of memory on searches over the 20MB output cap
  • Fixed a restored memory file's age note changing between requests after a compaction or resume, causing prompt cache misses
  • Fixed --resume and --continue dropping earlier thinking when a built-in tool the session started with has since been switched off server-side
  • Fixed /update-config writing Write(path) permission rules, which file permission checks don't match, instead of Edit(path) rules
  • Fixed SubagentStop hooks with a specific matcher firing for every stopping subagent whose agent type was empty
  • Fixed a terminal API Error: 400 on every turn behind a network gateway that rewrites API error responses when a beta request header is rejected
  • Fixed a crash at launch when ~/.claude.json holds a malformed mcpNeedsAuthNoticed value

2.1.276

  • Bug fixes and reliability improvements

Claude Code changes

New Documents

claude-projects [Source]

Projects, in public beta on Pro and Max, rolling out gradually starting with accounts that have used cloud sessions and don't already have projects in claude.ai chat or Cowork — not on Team or Enterprise yet, with a waitlist if the rollout hasn't reached you.

A project is one ongoing conversation in which Claude coordinates a stream of related work. Each task becomes a thread, and each thread is a cloud session running in parallel that keeps going after you close the laptop. The page frames it against doing the coordinating yourself:

  • Send work to one place: paste a bug report, stack trace or task list into the conversation, and Claude either starts a thread or passes it to the thread already working in that area, answering quick questions in place.
  • Set context once: every new thread starts with the project's repositories, instructions and memory, so a rule stated once — such as which branch to target — reaches all of them.
  • Come back to finished work: the Overview pane shows which threads finished, which pull requests are ready, and which thread is waiting on you.

Its twenty-odd sections cover creating a project from scratch or from an existing cloud session, GitHub access, reviewing a thread's pull request, opening a thread when you need direct control, choosing models and letting Claude manage context, unblocking a thread waiting on approval, writing project instructions, deciding which repositories to add, getting skills, plugins, connectors and tools into threads, and a full settings reference.

Two details worth flagging: project settings are changed at claude.ai/code or in the desktop app, not in settings.json; and changes to instructions, repositories, plugins and environment reach new threads only, not ones already running. Pause stops everything at once — running threads, the conversation, new threads and routines — until you resume.

Changed documents

agent-sdk/typescript [Source]

  • New startup_failure_reason field with a sixteen-value enum naming exactly why Claude Code refused to start — "so your application can offer the fix instead of a retry". The values run from org_pin_api_key_conflict and org_pin_mismatch through proxy_invalid, temp_dir_unusable, cwd_unavailable, shell_tool_missing, worktree_resume_refused and cli_version_too_old to bypass_root, each with a table row explaining what stopped the session. Requires Agent SDK v0.3.274. [lines 1416-1464] [Source]
  • By default the result is written only for three specific failures — the two worktree refusals and a refused continue of a background-held conversation. Set CLAUDE_CODE_STARTUP_FAILURE_RESULTS=1 to get it for every value; otherwise the rest end with stderr output, a non-zero exit, and no result message at all. [lines 1420-1423] [Source]
  • New mcpServer field on mcp__* tool entries naming the server that serves the tool and where its definition came from, plus a source field on each mcp_servers entry. Requires Agent SDK v0.3.274. [lines 995-1505] [Source]

discover-plugins [Source]

  • New "Add from claude.ai" section. In sessions where plugins sync from your claude.ai account, claude plugin marketplace list prints a From claude.ai: section, and claude plugin marketplace add --claudeai <name> adds one. Claude Code registers it under a local claudeai--prefixed name derived from the claude.ai listing — "Organization library" becomes claudeai-organization-library. Requires v2.1.273. [lines 274-284] [Source]
  • On sign-out or an account switch, such a marketplace stays configured but shows no plugins, while plugins you already installed from it keep loading. [line 286] [Source]
  • New "Add a marketplace and install in one command" section for --marketplace. The source takes the usual forms but can't contain spaces, the plugin name must be given bare without an @marketplace suffix, and Claude Code shows the resolved source and asks you to confirm before adding — declining cancels the install and adds nothing. Requires v2.1.275. [lines 334-344] [Source]
  • Marketplaces added from claude.ai get auto-update enabled by default, alongside claude-plugins-official and most official Anthropic marketplaces. [line 496] [Source]

env-vars [Source]

  • New CLAUDE_CODE_STARTUP_FAILURE_RESULTS, CLAUDE_CODE_MCP_STARTUP_WAIT_MS (v2.1.274), CLAUDE_CODE_AUTO_MODE_SERVER, CLAUDE_CODE_DISABLE_WINDOWS_SHELL_LAUNCHER, CLAUDE_CODE_GATEWAY_MODEL_DISCOVERY_TIMEOUT_MS and CLAUDE_CODE_WORKFLOW_MAX_CONCURRENT_AGENTS. [lines 206-388] [Source]
  • CLAUDE_CODE_BG_TASKS_REPORT_RUNNING documents a default flip: a non-interactive session now keeps reporting a running status past turn end while background work is live, so a host watching the status doesn't announce that Claude is waiting for input mid-work. Background shell commands don't hold it. Set 0 to opt out; on versions before v2.1.269 you had to set 1 to opt in. [line 209] [Source]
  • CLAUDE_CODE_WORKFLOW_MAX_CONCURRENT_AGENTS explains the memory tradeoff behind its default of 16: each running agent's transcript stays in Claude Code's memory. [line 388] [Source]
  • CLAUDE_CODE_RESUME_INTERRUPTED_TURN_MAX_AGE_MS gets a fuller entry: unset or 0 means no bound except that a turn whose last request failed with an API error resumes only while that error is under six hours old, a positive value bounds every turn, and a negative or non-numeric value applies a one-hour bound. [line 338] [Source]
  • MCP_CONNECTION_NONBLOCKING now spells out the exceptions to non-blocking startup: alwaysLoad: true servers still make startup wait (unless served from the discovery cache), and -p without --input-format stream-json waits for pending servers before the first turn regardless. [line 443] [Source]

errors [Source]

  • A large batch of new error entries, most of them authentication: Claude login not accepted for a cloud session the server refuses with a 401, four MCP re-authentication variants (including headersHelper and config-Authorization credential rejections and an insufficient-scope form), Google Cloud and Microsoft Foundry credential failures, Gateway refused the request, and Single sign-on authorization needed. [lines 65-189] [Source]
  • Two new output-style errors — custom styles can't be selected over Remote Control or from a relayed message, and styles saved to .claude/settings.local.json in a session that doesn't load it. [lines 197-198] [Source]
  • New symlink-race entries for Read and Grep: its symlink resolution changed after permission was checked. [line 225] [Source]

API changes

Changed documents

build-with-claude/compaction [Source]

  • The on-demand compaction section added two days earlier gained its remaining SDK language examples, including C#, alongside a note that a conversation with no system prompt or tools sends neither on the summarization request. max_tokens capping the whole call including thinking is now repeated inline in the examples. [lines 3427-3554] [Source]

build-with-claude/thinking [Source]

  • Streaming examples filled out across Python, Java and PHP, handling ThinkingDelta alongside text deltas and the content_block_start block types. [lines 563-728] [Source]
  • Claude Mythos Preview accepts max_tokens up to 128K, and the Batches beta ceiling isn't available for it. [line 1160] [Source]

managed-agents/events-and-streaming [Source] and managed-agents/migration

  • Both pages gained worked event-loop examples in more SDK languages, showing how to branch on ManagedAgentsAgentMessageEvent, ManagedAgentsAgentCustomToolUseEvent and ManagedAgentsSessionStatusIdleEvent, and how to detect an end-of-turn stop reason in each. [lines 1146-1238] [Source]

agents-and-tools/tool-use/fine-grained-tool-streaming [Source] and build-with-claude/handling-stop-reasons

  • Both gained PHP and Python examples, including distinguishing a max_tokens cutoff from natural completion by reading usage.output_tokens alongside the stop reason. [lines 3429-3434] [Source]

api/compliance [Source]

  • The compliance activity catalog expanded again across compliance.md, compliance/activities.md and the list endpoint, with new activity types and their object schemas.